Transformer substation fault handling method combining causal reasoning knowledge graph modeling

Abstract

The invention is suitable for the technical field of data analysis, and provides a transformer substation fault handling method combining causal reasoning knowledge graph modeling, comprising: acquiring multi-source heterogeneous data and performing data cleaning processing to obtain a space-time alignment data set, the space-time alignment data set comprising one or more quaternary data sets, the quaternary data set comprises a device identifier, a timestamp, a feature vector and an event tag; causal modeling processing is carried out on the time-space alignment data set to obtain a causal graph, and the causal graph comprises node information of nodes and relation information between the nodes; constructing a space-time diagram neural network model according to the causal diagram and the equipment connection relation diagram, wherein the space-time diagram neural network model realizes dynamic evolution of the graph based on an incremental updating strategy; and outputting fault root cause positioning information according to the time-space diagram neural network model.

Description

Technical Field

[0001] This application belongs to the field of data analysis technology, and in particular relates to a substation fault handling method that combines causal reasoning knowledge graph modeling. Background Technology

[0002] With the intelligent development of power systems, power plants, as key nodes in the transmission and transformation system, play a crucial role in the reliability of the entire power grid through their safe and stable operation. In recent years, with the increasing complexity of power plant equipment types and structures, traditional fault diagnosis methods relying on manual experience are no longer sufficient to meet the needs of intelligent operation and maintenance. Therefore, knowledge graph-based power plant fault modeling methods have gradually become a research hotspot.

[0003] Knowledge graphs, by constructing a data structure centered on "entity-relationship-attribute," can intuitively express the hierarchical structure, functional relationships, and causal paths of faults among devices. However, existing knowledge graph construction methods mostly focus on static data modeling, lacking the characterization of temporal evolution and dynamic changes in fault propagation paths. Furthermore, traditional methods often employ rule-based or data mining approaches for knowledge extraction and modeling, failing to fully incorporate causal reasoning theory and making it difficult to effectively identify the root causes of complex faults. Current technical solutions mainly fall into two categories: 1. Fault knowledge graphs based on expert rules and semantic ontology. This type of method relies on artificially constructed domain knowledge to build an ontology framework and uses expert experience rules for fault association modeling. Representative research includes a substation fault knowledge representation model based on OWL semantics (Web Ontology Language Semantics). The advantage of this method is that the model has strong interpretability, but it suffers from low modeling efficiency, poor adaptability, and difficulty in expansion. 2. Knowledge graph construction methods based on graph neural networks and data-driven approaches. With the development of machine learning, some studies have attempted to introduce techniques such as Graph Neural Networks (GNNs) to learn patterns in equipment faults and model the dependencies between devices using graph structures. These methods have improved the automation of modeling to some extent, but they usually rely on a large number of samples for training and lack explicit expression of causal mechanisms, resulting in insufficient generalization ability of the model when faced with novel fault scenarios and difficulty in accurately locating the root cause.

[0004] In summary, current knowledge graph modeling techniques for plant faults still face several major challenges in practical applications: 1. Lack of causal modeling mechanisms. Existing models mostly focus on association analysis, lacking causal chain modeling for fault occurrence, propagation, and spread, making it difficult to support root cause diagnosis and reasoning. 2. Static graph structure. Most knowledge graph construction methods are based on static topologies, failing to consider the dynamic changes in graph structure caused by the evolution of plant operating states over time, making it difficult to reflect the fault evolution path in actual operation. Therefore, a novel knowledge graph modeling method that integrates causal reasoning and dynamic graph structure evolution is urgently needed to improve the accuracy, dynamic adaptability, and reasoning ability of plant fault modeling, providing more reliable decision support for intelligent operation and maintenance systems. Summary of the Invention

[0005] This application provides a substation fault handling method that combines causal reasoning knowledge graph modeling, which can solve the technical problems of difficulty in performing fault root cause analysis and static knowledge graph modeling in the prior art.

[0006] In a first aspect, embodiments of this application provide a substation fault handling method combining causal reasoning knowledge graph modeling, including:

[0007] By collecting heterogeneous data from multiple sources and performing data cleaning, a spatiotemporally aligned dataset is obtained. The spatiotemporally aligned dataset includes one or more quadruple data sets, which include device identifier, timestamp, feature vector, and event tag.

[0008] Causal modeling is performed on the spatiotemporal aligned dataset to obtain a causal graph, which includes node information and relationship information between nodes.

[0009] A spatiotemporal graph neural network model is constructed based on the causal graph and the device connection graph. The spatiotemporal graph neural network model realizes dynamic evolution of the graph based on an incremental update strategy.

[0010] Based on the spatiotemporal graph neural network model, the fault root cause localization information is output.

[0011] In one possible implementation of the first aspect, the data cleaning process includes at least one of the following: invalid value handling, time alignment, feature normalization, and recoding.

[0012] In one possible implementation of the first aspect, the causal modeling process performed on the spatiotemporally aligned dataset to obtain a causal graph includes:

[0013] A sliding window-based Granger causality detection method is used to perform causal analysis on the device runtime sequence data in the spatiotemporally aligned dataset to obtain the first causal information;

[0014] The first causal information is corrected based on the device connection relationship graph to obtain the causal graph.

[0015] In one possible implementation of the first aspect, the node information of the node includes at least one of the following: device identifier, fault event, and status indicator; the relationship information between the nodes includes at least one of the following: causal dependency relationship between nodes, causal strength, time delay, and confidence label; the causal strength is determined based on Granger coefficient factor, weight coefficient of Granger coefficient factor, Bayesian confidence factor, weight coefficient of Bayesian confidence factor, historical fault co-occurrence frequency factor, and weight coefficient of historical fault co-occurrence frequency factor, wherein the weight coefficient of Granger coefficient factor, the weight coefficient of Bayesian confidence factor, and the weight coefficient of historical fault co-occurrence frequency factor are updated based on time changes or system state changes.

[0016] In one possible implementation of the first aspect, the construction of a spatiotemporal graph neural network model based on the causal graph and the device connection graph, wherein the spatiotemporal graph neural network model realizes dynamic evolution of the graph based on an incremental update strategy, includes:

[0017] A knowledge graph ontology structure is constructed based on the device connection relationship graph. The knowledge graph ontology structure includes entity type, relationship type, and attribute information.

[0018] Based on the knowledge graph ontology structure and the causal graph, a spatiotemporal graph neural network model is constructed. The spatiotemporal graph neural network model implements an update strategy to achieve dynamic evolution of the graph when a new alarm event is generated or when the causal weight changes exceed a threshold.

[0019] In one possible implementation of the first aspect, the step of outputting fault root cause localization information based on the spatiotemporal graph neural network model includes at least one of the following:

[0020] The active alarm list is input into the spatiotemporal graph neural network model, which outputs multiple device identifiers as fault root cause location information.

[0021] The simulated scenario is input into the spatiotemporal graph neural network model, and a quantitative comparison report is output as the fault root cause location information. The simulated scenario includes any one of the following: simulated parameter correction scenario, simulated topology reconstruction scenario, simulated environmental intervention scenario, simulated protection logic optimization scenario, and simulated operation and maintenance measurement adjustment scenario.

[0022] When abnormal parameters are detected, warning information is output based on the spatiotemporal graph neural network model.

[0023] In one possible implementation of the first aspect, outputting fault root cause localization information based on the spatiotemporal graph neural network model includes:

[0024] The causal topology of the knowledge graph and real-time device data are input into the spatiotemporal graph neural network model to obtain multiple risk paths, which include paths between multiple nodes.

[0025] A random walk algorithm with restart is used to evaluate the risk coefficients of the multiple risky paths and identify one or more high-risk paths.

[0026] The simulated scenario corresponding to one or more high-risk paths is input into the spatiotemporal graph neural network model, and the key causal path is output as the root cause location information of the fault and the spatiotemporal graph neural network model is optimized.

[0027] Secondly, embodiments of this application provide a substation fault handling device that combines causal reasoning knowledge graph modeling, including:

[0028] The first acquisition module is used to acquire spatiotemporally aligned datasets by acquiring multi-source heterogeneous data and performing data cleaning processing. The spatiotemporally aligned datasets include one or more four-element data groups, which include device identifier, timestamp, feature vector, and event tag.

[0029] The second modeling module is used to perform causal modeling processing on the spatiotemporal aligned dataset to obtain a causal graph, which includes node information and relationship information between nodes.

[0030] The third construction module is used to construct a spatiotemporal graph neural network model based on the causal graph and the device connection graph. The spatiotemporal graph neural network model realizes dynamic evolution of the graph based on an incremental update strategy.

[0031] The fourth service module is used to output fault root cause location information based on the spatiotemporal graph neural network model.

[0032] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the electronic device performs the method as described in any one of the first aspects above.

[0033] Fourthly, embodiments of this application provide a computer program product, including a computer program, which, when run, causes the method as described in any one of the first aspects above to be performed.

[0034] Fifthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as described in any one of the first aspects above.

[0035] The beneficial effects of the first aspect of this application compared with the prior art are:

[0036] This application, based on traditional knowledge graph construction, adds a causal reasoning step. By explicitly modeling the causal chain of equipment state changes, the root cause of faults can be accurately identified. This application overcomes the shortcomings of existing technologies based on correlation analysis, not only discovering connections between fault phenomena but also deeply analyzing the causal paths of faults, significantly improving the accuracy and interpretability of fault diagnosis. By introducing a graph structure evolution mechanism, this application automatically updates the knowledge graph structure based on real-time changes in plant operating status and the dynamic process of fault propagation. Graph structure evolution makes the knowledge graph more adaptable to complex fault scenarios, promptly reflecting the impact of factors such as fault propagation, equipment maintenance, and replacement on equipment relationships, improving the system's timeliness and accuracy in responding to plant state changes. By integrating causal reasoning and graph reasoning mechanisms, this application not only supports fault root cause localization based on current data but also performs complex multi-level fault prediction and propagation path reasoning. This application effectively supports intelligent decision-making assistance in operation and maintenance decision-making systems, enabling early identification of potential fault risks and emergency response. For example, the system can predict possible fault propagation paths based on the causal relationships of historical events and provide early warnings and operational suggestions to maintenance personnel. This application's embodiments introduce intelligent and visual steps into the knowledge graph construction process. Through the visual graph interface, maintenance personnel can intuitively see the causal relationships between devices and the fault evolution process. The graph not only has clear hierarchy and intuitive structure, but also dynamically displays the evolution of plant equipment status over time, providing maintenance personnel with real-time and accurate fault diagnosis basis, significantly improving fault analysis efficiency.

[0037] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart illustrating a substation fault handling method based on causal reasoning knowledge graph modeling, provided in one embodiment of this application.

[0040] Figure 2 This is a flowchart illustrating the output of fault root cause location information provided in an application embodiment of this application;

[0041] Figure 3 This is a flowchart illustrating the output of fault root cause location information provided in another application embodiment of this application;

[0042] Figure 4 This is a flowchart illustrating a substation fault handling method based on causal reasoning knowledge graph modeling, provided in an application embodiment of this application.

[0043] Figure 5 This is a schematic diagram of the structure of a substation fault handling device that combines causal reasoning knowledge graph modeling, as provided in an embodiment of this application.

[0044] Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0045] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0046] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0047] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0048] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0049] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0050] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0051] Figure 1 This is a flowchart illustrating a substation fault handling method based on causal reasoning knowledge graph modeling, provided in one embodiment of this application.

[0052] S11. By collecting multi-source heterogeneous data and performing data cleaning processing, a spatiotemporal aligned dataset is obtained. The spatiotemporal aligned dataset includes one or more four-element data groups, which include device identifier, timestamp, feature vector, and event label.

[0053] The multi-source heterogeneous data includes, but is not limited to, SCADA (Supervisory Control and Data Acquisition, or SCADA for short, real-time or historical data collected, processed and managed by monitoring and data acquisition systems), secondary equipment logs, equipment ledgers, historical fault reports, etc.

[0054] This application embodiment can access multi-source heterogeneous data such as SCADA real-time data, secondary equipment logs, equipment ledgers, and historical fault reports. Data cleaning processing is performed on this multi-source heterogeneous data to obtain a spatiotemporally aligned dataset. Specifically, data cleaning processing of multi-source heterogeneous data can be performed in the following ways: For example, for data preprocessing, an algorithm based on DTW (Dynamic Time Warping) can be used to solve the problem of inconsistent sampling frequencies between devices. Simultaneously, a rule engine is used to perform logical verification on the logs, eliminating conflicting and abnormal records. In terms of feature engineering, power industry feature templates such as "voltage descent rate" and "insulation aging index" are defined, and time-series features are extracted using a sliding window. For text-based data, the BERT+BiLSTM model (Bidirectional Encoder Representations from Transformers + Bidirectional Long Short-Term Memory, where Bidirectional Encoder Representations from Transformers refers to a pre-trained language model based on Transformer encoders that can understand text semantics through bidirectional context, and Bidirectional Long Short-Term Memory refers to a bidirectional long short-term memory network that can capture forward and backward temporal dependencies in sequence data, abbreviated as BERT+BiLSTM model) is introduced to extract entity relations, transforming multi-source heterogeneous data into structured four-element data sets. The format of the four-element data set is: "Device Identifier (Device ID), Timestamp, Feature Vector, Event Label". One or more four-element data sets constitute a spatiotemporally aligned dataset.

[0055] Let's illustrate this with an example. The monitoring system detects a sudden drop in the 220kV bus voltage to 0, accompanied by the tripping of three circuit breakers. In this scenario, the collected multi-source heterogeneous data includes raw data 1, raw data 2, and raw data 3. Raw data 1 is SCADA data (1 record per second), specifically: "{timestamp: 2023-05-10 14:05:32, device ID: BUS-220, voltage: 0kV, current: 0A}". Raw data 2 is the protection device log, specifically: "[14:05:30] Line protection L1 started → [14:05:31] Circuit breaker CB1 / CB2 / CB3 tripped". Raw data 3 is the equipment ledger, specifically: "Bus BUS-220 connects to CB1 / CB2 / CB3, downstream to transformer T1". By collecting heterogeneous data from multiple sources and performing data cleaning, a spatiotemporally aligned dataset can be obtained. The spatiotemporally aligned dataset includes the following three quadruple data sets, as shown in Table 1.

[0056] Table 1, Spatiotemporal Alignment Dataset 1.

[0057]

[0058] S12, perform causal modeling processing on the spatiotemporal aligned dataset to obtain a causal graph, which includes node information and relationship information between nodes.

[0059] In the causal modeling stage, Granger causal analysis can be used to identify lagged correlations between equipment state variables, forming preliminary causal hypotheses. Next, a prior Bayesian network is constructed based on the equipment's physical topology (e.g., the connection between circuit breakers and busbars, and other connection relationships) to constrain the causal search space. Subsequently, a causal discovery algorithm based on independence testing is used to perform conditional independence checks, eliminating false causal relationships and obtaining the final set of causal edges, which constitutes a causal graph. The causal graph includes node information and information about the relationships between nodes, which can be represented by edges between nodes. For example, node information includes, but is not limited to, equipment identification, fault events, and state indicators. The information about the relationships between nodes (edges between nodes) includes, but is not limited to, causal dependencies, causal strength, time delays, and confidence labels.

[0060] Following the previous example, causal modeling can be performed on the spatiotemporally aligned dataset and the historical fault database (e.g., including fault data from the past 3 years). First, Granger causal analysis reveals that "transformer T1 oil temperature rises (starting at 14:05:00) → bus voltage drops sharply (lag 32 seconds, F-statistic is 5.2, p-value calculated by the random walk algorithm with restart is 0.003)". Then, Bayesian correction is performed, increasing the confidence level to 0.85 based on the electrical connection diagram ("T1→BUS-220" physical direct connection). Co-occurrence analysis then reveals that "T1 overload and bus power failure co-occurred 7 times historically (TF-IDF weight 0.73)". The causal graph obtained through causal modeling is shown below.

[0061] Node: [T1]--(CAUSES,W=0.89)--> Node: [BUS-220]--(TRIGGERS,W=0.75)--> Node: [CB1 Trip]. Here, "[T1]", "[BUS-220]", and "[CB1 Trip]" represent node information. "(CAUSES,W=0.89)" and "(TRIGGERS,W=0.75)" represent relationship information between nodes. "CAUSES" indicates a causal relationship between nodes, "TRIGGERS" indicates a triggering relationship, and "ELECTRICALLY_CONNECTED" indicates an electrical connection. W represents the causal strength.

[0062] S13, a spatiotemporal graph neural network model is constructed based on the causal graph and the device connection relationship graph. The spatiotemporal graph neural network model realizes dynamic evolution of the graph based on an incremental update strategy.

[0063] In terms of graph construction, a hierarchical ontology design approach can be adopted, dividing the spatiotemporal graph neural network into a physical layer, a functional layer, and a causal layer. The physical and functional layers are constructed using a device connection graph. The physical layer carries device connection relationships (e.g., it can inherit the CIM model structure, Common Information Model), while the functional layer represents power transmission logic (e.g., "transformer carrying load"). The causal layer is constructed using a causal graph. The causal layer describes the fault propagation path generated by the causal graph. Furthermore, a spatio-temporal graph neural network (STGNN) can be introduced to model the dynamic evolution of the graph. The TGAT model (Temporal Graph Attention Network, TGAT) encodes the temporal features of nodes, aggregating the states of neighboring nodes through graph convolution to achieve dynamic updates of edge weights. Regarding graph evolution, the spatio-temporal graph neural network model can achieve dynamic graph evolution based on an incremental update strategy. When a new fault event is detected, only its k hop neighborhood subgraphs can be locally updated. During off-peak hours each day, full graph sparsity optimization can be performed to remove edges with a confidence level below 0.2.

[0064] Following the previous example, after inputting the causal graph and the device ontology graph (the standard for CIM models), the graph construction process includes ontology layering and dynamic updates via a spatiotemporal graph neural network. The ontology layering results are, for example, "Physical layer: Transformer T1 -- ELECT_CONNECTED --> Busbar BUS-220", and "Causal layer: T1 -- CAUSES --> BUS-220 (weight 0.89, last update time 14:10:00)". The spatiotemporal graph neural network dynamic updates are: "Add T1 overload event → Reconstruct subgraph (3-hop neighborhood of T1)", and update edge weights "CAUSES edge weight from 0.89 → 0.92 (due to continuous oil temperature rise)". For example, Neo4j can be used to store the graph, with the graph storage structure "(:Transformer{id:"T1",temp:72})-[:CAUSES{weight:0.92,confidence:0.93}]->(:Busbar{id:"BUS-220",voltage:0})".

[0065] S14, based on the spatiotemporal graph neural network model, output the root cause location information of the fault.

[0066] Based on the spatiotemporal graph neural network model, fault root cause location information can be output in various ways. Here, information such as the currently active alarm list and real-time alarm information can be sent as input to the spatiotemporal graph neural network model to obtain the fault root cause location information output by the model. Alternatively, the spatiotemporal graph neural network model can automatically process and analyze the information without providing any input, outputting the fault root cause location information. Those skilled in the art should understand that both application scenarios with and without input information fall within the scope of this application.

[0067] In one embodiment, the data cleaning process described in step S11 of the above embodiment includes at least one of the following: invalid value processing, time alignment, feature normalization, and recoding.

[0068] Multi-source heterogeneous data includes, but is not limited to, SCADA data, secondary equipment logs, equipment ledger data, and historical fault reports. SCADA data is acquired using the IEC 61850 protocol to collect real-time substation telemetry (e.g., voltage, current), telesignal (e.g., circuit breaker status), and equipment operating status data. Secondary equipment logs are acquired using the syslog and SNMP (Simple Network Management Protocol) protocols to obtain the operating logs and system logs of secondary equipment. Equipment ledger data is generated by automatically constructing the substation topology and equipment ledger based on the equipment model described in the IEC 61850 protocol and network traffic analysis, ensuring that the equipment connection relationships are consistent with the actual physical wiring relationships. Historical fault reports are obtained by retrieving fault recording files and historical data from protection devices using the IEC 61850 protocol for subsequent causal analysis and model training.

[0069] Data cleaning methods for multi-source heterogeneous data include, but are not limited to, at least one of the following: invalid value handling, time alignment, feature normalization, and recoding.

[0070] The invalid value handling method is to mark data that exceeds the rated range of the equipment for N consecutive sampling points (e.g., 3) as "out-of-limit data" or "invalid data", and fill it with linear interpolation of the previous valid value.

[0071] The time alignment method involves synchronizing the clocks of all devices to ensure that the data timestamp error is less than the error time threshold (e.g., less than 1ms). For devices with large clock deviations, a buffer queue is used for data realignment.

[0072] The feature normalization method is to convert analog quantities (such as voltage and current) into per-unit values.

[0073] The recoding method involves encoding discrete state quantities (such as the opening and closing of circuit breakers) into binary vectors.

[0074] By collecting heterogeneous data from multiple sources and performing data cleaning, a standardized and structured spatiotemporally aligned dataset can be obtained, which can be stored in time series databases (InfluxDB) and graph databases (such as Neo4j) for subsequent analysis.

[0075] In one embodiment, step S12 in the above embodiment, which involves performing causal modeling on the spatiotemporally aligned dataset to obtain a causal graph, includes the following steps S121 and S122.

[0076] S121, the Granger causal detection method based on sliding window is used to perform causal analysis on the device runtime sequence data in the spatiotemporal aligned dataset to obtain the first causal information.

[0077] For example, an analysis strategy with a first predetermined time window (e.g., 60 seconds) and a second predetermined time sliding step (e.g., 10 seconds) can be used to perform causal analysis on the time series data of devices in a spatiotemporally aligned dataset. For instance, firstly, an ADF (Augmented Dickey-Fuller Test) stationarity test (significance level, e.g., p < 0.05) is performed on the time series data of devices A and B within the window. Data that pass the stationarity test are then subjected to a Granger causality test. When the calculated F-statistic is greater than 4.0 and the p-value is less than 0.01, it is determined that device A has a causal relationship with device B. All causal relationships that pass the test form an initial set of causal edges, obtaining the first causal information, where each edge records the corresponding F-statistic value.

[0078] S122, the first causal information is corrected based on the device connection relationship diagram to obtain a causal diagram.

[0079] Connection relationships can be obtained from a device connection graph (e.g., a CIM model) and converted into an adjacency matrix Adj_matrix, where an element Adj_matrix[A][B] = 1 indicates that devices A and B are physically connected. The initial causal information obtained from Granger causality analysis is then corrected by deleting causal edges where Adj_matrix[A][B] = 0 (physically unreachable) and adding causal edges where Adj_matrix[A][B] = 1 but missed by Granger analysis (the default weight can be, for example, 0.5), thus obtaining the causal graph.

[0080] In one embodiment, the node information of the node includes at least one of the following: device identifier, fault event, and status indicator; the relationship information between the nodes includes at least one of the following: causal dependency relationship between nodes, causal strength, time delay, and confidence label; the causal strength is determined based on Granger coefficient factor, weight coefficient of Granger coefficient factor, Bayesian confidence factor, weight coefficient of Bayesian confidence factor, historical fault co-occurrence frequency factor, and weight coefficient of historical fault co-occurrence frequency factor, wherein the weight coefficient of Granger coefficient factor, the weight coefficient of Bayesian confidence factor, and the weight coefficient of historical fault co-occurrence frequency factor are updated based on time changes or system state changes.

[0081] In the causal reinforcement stage, a dynamic causal weighting mechanism may be involved, and the causal strength can be calculated in the following formula (1).

[0082] W = α{Granger coefficient factor} + β{Bayes confidence factor} + γ{historical fault co-occurrence frequency factor}

[0083] Calculation formula (1)

[0084] Where W represents causal strength, α represents the weighting coefficient of the Granger coefficient factor, β represents the weighting coefficient of the Bayesian confidence factor, and γ represents the weighting coefficient of the historical fault co-occurrence frequency factor. The weighting coefficients of the Granger coefficient factor, the Bayesian confidence factor, and the historical fault co-occurrence frequency factor can be adaptively adjusted according to the equipment's operating status. A weighted approach can be used to calculate the causal strength of the causal edge, with initialization parameters such as α = 0.6, β = 0.3, and γ = 0.1.

[0085] Granger coefficients can be obtained by quantifying the intensity of fault propagation between devices through time-series causal analysis. First, second-level monitoring data (such as voltage and current) of the target device is collected. After stationarity testing and standardization preprocessing, the data is segmented using a sliding window mechanism (e.g., a 60-second window with a 10-second step). Within the window, a Vector Autoregression (VAR) model is constructed, and two prediction models for device B are compared: the baseline model uses only historical data from device B, while the extended model incorporates historical data from device A. Hypothesis testing (F-statistic) is used to determine whether the prediction accuracy of the extended model is significantly improved. When the p-value is less than 0.01, Granger coefficients are generated. After normalization, the Granger coefficients range from 0 to 1; a larger value indicates a stronger causal influence of device A on device B.

[0086] The Bayesian confidence factor integrates physical connections between devices with historical failure probabilities to assess the reliability of causal relationships. A prior topology network is constructed based on the main electrical wiring diagram, assigning an initial weight of 0.6 to edges of directly physically connected devices. Based on device operation records and historical failure logs, the conditional probability of device B failing after device A fails is calculated, and the posterior probability is derived using Bayes' theorem. The final Bayesian confidence factor is the ratio of the posterior probability to the prior probability, with adjustments made based on the physical distance between devices.

[0087] The historical fault co-occurrence frequency factor mines implicit correlations from long-term operation and maintenance data. It filters protection action-related fault records from the past three years, defining a co-occurrence time window for fault events, for example, ±15 seconds. It counts the number of times device B fails within the time window after device A fails, and then uses a TF-IDF algorithm for weighted processing: the numerator is the co-occurrence frequency, and the denominator incorporates the number of independent faults of device A for logarithmic decay, eliminating statistical interference from high-frequency fault devices. Simultaneously, an exponential decay model is used to update the data monthly (attenuation factor 0.95) to ensure timeliness. The historical fault co-occurrence frequency factor is particularly suitable for identifying implicit correlations that are not physically connected (such as cases where lightning strikes cause multiple devices to fail simultaneously), and its output value is normalized and mapped to the 0-1 interval.

[0088] The weight coefficients of the Granger coefficient factor, the Bayesian confidence factor, and the historical fault co-occurrence frequency factor can be updated based on time changes or system state changes. For example, the system automatically recalculates the weights every hour, or triggers the knowledge graph update process when the weight change exceeds a predetermined threshold (e.g., 15%).

[0089] In one embodiment, step S13 in the above embodiment, which involves constructing a spatiotemporal graph neural network model based on the causal graph and the device connection graph, wherein the spatiotemporal graph neural network model achieves dynamic evolution of the graph based on an incremental update strategy, includes the following steps S131 and S132.

[0090] S131, Construct a knowledge graph ontology structure based on the device connection relationship graph, wherein the knowledge graph ontology structure includes entity type, relationship type, and attribute information.

[0091] A knowledge graph ontology structure can be defined based on an equipment connection graph conforming to the power system CIM standard. This ontology structure includes the following core elements: entity type, relationship type, and attribute information. Entity types include, but are not limited to, power equipment entities such as busbars, circuit breakers, transformers, and protection devices. Relationship types include, but are not limited to, three core types: electrical connection relationships, causal relationships, and triggering relationships. Attribute information includes, but is not limited to, equipment node attributes and causal edge attributes. Equipment node attributes include rated voltage, installation date, and manufacturer. Causal edge attributes include, but are not limited to, confidence score and last updated time.

[0092] S132, Based on the knowledge graph ontology structure and the causal graph, a spatiotemporal graph neural network model is constructed. The spatiotemporal graph neural network model implements an update strategy to achieve dynamic evolution of the graph when a new alarm event is generated or when the causal weight change exceeds a threshold.

[0093] A spatiotemporal graph neural network model can be constructed based on the knowledge graph ontology structure and causal graph to process spatiotemporal features. Input features include node feature vectors and edge feature vectors. Node feature vectors include, but are not limited to, per-unit voltage (V_pu), current (I_pu), and device status code (status_code); edge feature vectors include, but are not limited to, dynamic causal weights (W_ABt) and physical distance (physical_distance). The network architecture of the spatiotemporal graph neural network model can include temporal convolutional layers and graph attention layers. The temporal convolutional layer can, for example, use a 3-layer TCN structure with a sliding window size of 5 to extract temporal features. The graph attention layer can, for example, use a 2-layer GAT structure with 4 heads to capture spatial relationships. Training data can, for example, use historical data from the past 30 days, sampled every 5 minutes, for a total of 8640 training samples.

[0094] The spatiotemporal graph neural network model enables incremental updates to knowledge graphs. Update triggering conditions include two scenarios: Condition 1 is the generation of a new alarm event, such as a circuit breaker tripping, which automatically locates the alarm device and performs a local update of the subgraph within its three-hop range. Condition 2 is the detection of a causal weight change exceeding a threshold (15%), triggering feature updates for associated edges. The update operation can, for example, include three steps: 1. Recalculating node embedding vectors for the affected subgraph. 2. Fine-tuning the STGNN model using a predetermined learning rate (e.g., 0.001). 3. Generating a new graph version number, such as KG_v2.3.1, while retaining historical versions for backtracking analysis. Performance optimization employs subgraph segmentation techniques, updating only the affected regions to reduce computational overhead.

[0095] In one embodiment, step S14 in the above embodiment, which involves outputting fault root cause localization information based on the spatiotemporal graph neural network model, includes at least one of the following S141, S142, and S143.

[0096] S141, input the active alarm list into the spatiotemporal graph neural network model, and output multiple device identifiers as fault root cause location information.

[0097] For example, a fault root cause path analysis report can be output based on a multipath propagation algorithm. Here, the currently active alarm list can be sent as input to the spatiotemporal graph neural network model. This list can contain information about all devices that triggered the alarms. The spatiotemporal graph neural network model can trace back along the causal relationship edges (also known as CAUSES relationship edges) from each alarm node, with a maximum tracing depth, for example, set to 3 hops. The spatiotemporal graph neural network model can use a "random walk with restart" algorithm to score each tracing path, ultimately outputting, for example, the top 3 devices with the highest scores as root cause candidates, along with a complete fault root cause path analysis report, which includes a detailed scoring calculation process for each path.

[0098] S142, input the simulated scenario into the spatiotemporal graph neural network model and output a quantitative comparison report as fault root cause location information. The simulated scenario includes any one of the following: simulated parameter correction scenario, simulated topology reconstruction scenario, simulated environmental intervention scenario, simulated protection logic optimization scenario, and simulated operation and maintenance measurement adjustment scenario.

[0099] An interactive interface can be provided to receive simulated scenarios input by users and input the simulated scenarios into the spatiotemporal graph neural network model. After the spatiotemporal graph neural network model outputs a quantitative comparison report, the quantitative comparison report is displayed through the interactive interface as fault root cause location information.

[0100] After receiving a simulated scenario, the spatiotemporal graph neural network model can temporarily remove specified device nodes and all their associated edges from the knowledge graph. Then, based on the fundamental laws of the power system, it recalculates the electrical parameters of the affected equipment, including but not limited to changes in key indicators such as voltage and current. Finally, it generates a quantitative comparison report, which can display the magnitude of the change in fault probability under the simulated scenario as a percentage, as well as the specific numerical changes in key parameters, such as "bus power failure probability decreased by 82%, CB1 trip probability decreased by 79%".

[0101] Step S142 can also be called the counterfactual reasoning step. Common simulation scenarios include at least one of the following: Simulating parameter correction scenarios, assuming the equipment threshold is adjusted (e.g., reducing the transformer oil temperature alarm value from 75℃ to 70℃), and simulating whether the fault is avoided. Simulating topology reconfiguration scenarios, assuming the addition of protection devices on the critical path (e.g., busbar sectionalizing circuit breakers), and verifying whether fault propagation is blocked. Simulating environmental intervention scenarios, assuming the improvement of external conditions (e.g., reducing the grounding resistance from 5Ω to 1Ω during a lightning strike), and analyzing whether the fault current is effectively diverted. Simulating protection logic optimization scenarios, assuming the shortening of protection action time (e.g., increasing line protection speed from 0.5s to 0.2s), and verifying whether fault propagation is prevented. Simulating maintenance strategy adjustment scenarios, assuming the early implementation of preventative measures (e.g., activating backup cooling fans 1 hour before a fault), and assessing whether the oil temperature rise trend can be suppressed. The purpose of the counterfactual reasoning step is to transform passive fault handling into an active defense closed loop, achieving dynamic optimization of maintenance strategies and continuous improvement of system reliability through the four steps of "simulation-verification-execution-evolution".

[0102] S143, When an abnormal parameter is detected, an early warning message is output based on the spatiotemporal graph neural network model.

[0103] For example, the system can continuously monitor changes in key indicators within the knowledge graph. When the following abnormal parameters are detected, an early warning mechanism is triggered, and warning information is output based on the spatiotemporal graph neural network model. For instance, causal weights fluctuate beyond a predetermined threshold (e.g., 30%) within a third predetermined time period (e.g., 1 hour). Another example is that device status parameters continuously deviate from the normal range for more than a fourth predetermined time period (e.g., 5 minutes). Yet another example is that newly discovered causal relationships significantly differ from historical patterns. Warning information can be displayed in real-time through a visual interface and simultaneously pushed to the terminals of relevant maintenance personnel.

[0104] Figure 2 This is a flowchart illustrating the output of fault root cause location information provided in an application embodiment of this application.

[0105] like Figure 2 As shown, the following three services can be provided at the service layer based on the spatiotemporal graph neural network model.

[0106] In the root cause localization service, the list of active alarms can be input into the spatiotemporal graph neural network model, and multiple device identifiers can be output as fault root cause localization information by restarting the random walk algorithm (RWR).

[0107] In the counterfactual reasoning service, a simulated scenario can be input into the spatiotemporal graph neural network model, and a quantitative comparison report can be output as fault root cause location information. The fault root cause location information can be presented in a visual way through GNNExplainer (a tool for interpreting graph neural network prediction results).

[0108] In the prediction and early warning service, when abnormal parameters are detected, risk prediction can be performed based on the LSTM model in the spatiotemporal graph neural network model, and early warning information can be output.

[0109] The visualization interface can display fault root cause location information in various ways, such as fault propagation animation, evidence chain topology diagram, confidence heat map, etc.

[0110] In one embodiment, step S14 in the above embodiment, which involves outputting fault root cause location information based on the spatiotemporal graph neural network model, includes the following steps S144 to S146.

[0111] S144, the causal topology of the knowledge graph and real-time device data are input into the spatiotemporal graph neural network model to obtain multiple risk paths, which include paths between multiple nodes.

[0112] Here, early warning and proactive defense against faults can be achieved based on the causal topology of the knowledge graph and real-time device data. First, dynamic risk features are extracted from the knowledge graph, specifically including: reflecting system vulnerability through the weight change rate of key causal paths, and generating composite risk indicators by combining the changing trends of associated device states. Simultaneously, alarm density within the device's neighborhood is statistically analyzed to capture local anomaly clustering signals. These risk features, together with real-time time-series data of the devices (such as second-level monitoring values ​​of voltage and temperature), constitute the input matrix for a fifth predetermined time (e.g., 60 minutes) sliding window. A spatiotemporal graph neural network model combining bidirectional LSTM (Long Short-Term Memory) with an attention mechanism can be used to process these risk features. The LSTM layers learn the long-term dependencies of device state evolution, and the attention mechanism automatically focuses on abrupt changes and high-weight causal paths. The spatiotemporal graph neural network model can output multiple risk paths, each including paths between multiple nodes.

[0113] S145, using a random walk algorithm with restart, evaluate the risk coefficients of the multiple risky paths and determine one or more high-risk paths.

[0114] During root cause analysis, a random walk algorithm with restart (or random walk with restart) can be used to evaluate the risk coefficients of the multiple risk paths. For example, the risk coefficients of risk nodes in each risk path can be calculated using the following formula (2), and the risk coefficients of the risk paths can be determined based on the risk coefficients of each risk node.

[0115]

[0116] v i v j Indicates a risk node. P(v) i ) represents the risk coefficient of a risk node. A represents the set of alarm nodes. W ji Indicates v j to v i The weight of the edges between them. D(v) j ) represents node v j The out-degree or normalization factor. L represents the out-degree from node v. j to v i The path length. λ represents the path attenuation coefficient, which can be set according to the actual scenario; the default value can be, for example, 0.5.

[0117] After determining the risk coefficient of each risk path, you can select risk paths with risk coefficients greater than the risk threshold, or select risk paths with risk coefficients ranked in the top K (Top-K) as high-risk paths. High-risk paths can include a list of root cause devices and the corresponding chain of evidence.

[0118] S146, input the simulated scenario corresponding to one or more high-risk paths into the spatiotemporal graph neural network model, output the key causal path as the root cause location information of the fault, and optimize the spatiotemporal graph neural network model.

[0119] In counterfactual reasoning, simulated scenarios corresponding to one or more high-risk paths can be input into the spatiotemporal graph neural network model, and the key causal paths can be output as fault root cause localization information. For example, the impact of specific operation and maintenance (such as disconnecting a circuit breaker) on fault propagation can be simulated to generate a simulation scenario report, and the key causal paths as fault root cause localization information can be visualized through the GNN Explainer.

[0120] The above steps can also be used to optimize the spatiotemporal neural network model. For example, if the spatiotemporal neural network model detects that "T1 oil temperature rises at a rate of 1℃ / min", and through root cause analysis and counterfactual reasoning, it predicts that "the probability of failure within 2 hours is 0.83", then the spatiotemporal neural network model can automatically strengthen the monitoring of the high-risk path "T1→BUS-220", for example, by increasing the data sampling frequency of this high-risk path, such as to once per second.

[0121] Figure 3 This is a flowchart illustrating the output of root cause location information for faults, provided in another application embodiment of this application.

[0122] like Figure 3 As shown, the LSTM risk prediction step can achieve early warning. The causal topology of the knowledge graph and real-time device data are input into the spatiotemporal graph neural network model to obtain multiple risk paths, including paths between multiple nodes. The real-time root cause localization step can use a random walk algorithm with restart to evaluate the risk coefficients of the multiple risk paths and identify one or more high-risk paths. The counterfactual simulation step can verify the prediction. The simulated scenarios corresponding to the one or more high-risk paths are input into the spatiotemporal graph neural network model, outputting key causal paths as fault root cause localization information and optimizing the spatiotemporal graph neural network model.

[0123] Figure 4 This is a flowchart illustrating a substation fault handling method that combines causal reasoning knowledge graph modeling, as provided in an application embodiment of this application.

[0124] like Figure 4 As shown in the embodiments of this application, a fault root cause analysis platform for intelligent operation and maintenance of power systems is provided. By integrating dynamic causal reasoning and spatiotemporal knowledge graph technology, it can achieve accurate location, propagation and inference of equipment faults and prevention and early warning.

[0125] The core functional modules may include a real-time data hub, a causal modeling engine, a dynamic knowledge graph, and intelligent diagnostic services.

[0126] At the data layer, the real-time data hub can integrate multi-source information such as SCADA real-time monitoring data, secondary equipment logs, equipment ledgers, and historical fault data, and construct a unified spatiotemporal aligned dataset through spatiotemporal alignment technology.

[0127] At the modeling layer, the causal modeling engine can use a hybrid reasoning method that combines Granger causal analysis with Bayesian network correction to generate weighted causal graphs.

[0128] At the evolutionary level, dynamic knowledge graphs can construct spatiotemporal knowledge graphs based on the STGNN model and support a k-hop neighborhood incremental update mechanism.

[0129] At the service layer, intelligent diagnostic services can provide core functions such as fault root cause location services, counterfactual reasoning services, and risk prediction and early warning services, and intuitively display fault propagation paths and evidence chains through a visual interface.

[0130] The embodiments of this application can be mainly applied to two application scenarios: power system operation and maintenance and fault diagnosis, and intelligent dispatching and power grid security defense.

[0131] When applied to power system operation and maintenance and fault diagnosis scenarios, the technical solution of this application embodiment can be deployed in the intelligent operation and maintenance system of power plants and substations, accessing real-time SCADA data and protection equipment data, and realizing equipment anomaly detection, alarm attribution, fault path tracing, and emergency suggestions through dynamic knowledge graph construction. Typical applications may include: root cause location of abnormal transformer temperature rise; causal chain reconstruction of bus power failure events; and visualization of the multi-device linkage fault propagation process.

[0132] When applied to intelligent dispatching and power grid security defense scenarios, the application can be carried out by integrating the technical solution of this application embodiment with the power grid dispatching platform. During the execution of dispatching instructions, the risks that may be caused by changes in equipment status can be identified in advance based on the graph reasoning results, thus assisting in the formulation of dispatching strategies. Typical applications may include: risk prediction before high-voltage line switching; early warning of cascading faults after changes in critical loads; and identification of power grid vulnerabilities during extreme weather.

[0133] In the existing technology, there are three main technical problems in the modeling of knowledge graphs for plant faults.

[0134] 1. Lack of causal reasoning ability, making it difficult to effectively identify the root cause of faults. Existing technical solutions are mainly based on correlation analysis, which makes it difficult to establish causal chains between changes in equipment state. This results in the inability to accurately infer the source and propagation path of faults in multi-source and multi-level fault scenarios, limiting the accuracy and effectiveness of fault diagnosis. The embodiments of this application adopt a hybrid causal reasoning engine: integrating Granger causal analysis, Bayesian networks, and counterfactual reasoning to construct a multimodal causal discovery mechanism.

[0135] 2. Static graph structures are difficult to reflect the fault evolution process. Most current knowledge graphs use static graph structures, which cannot dynamically reflect the changes in plant equipment status over time. Especially in multi-step, multi-stage fault processes, they cannot update node attributes and edge dependencies, affecting the timeliness and expressive power of the knowledge graph. This application's embodiment employs spatiotemporal-aware graph evolution: a dynamic update algorithm based on the spatiotemporal graph neural network model (STGNN) supports dynamic reconstruction of the graph structure and adaptive weight updates in the time dimension, accurately modeling the evolution process of equipment status and fault propagation paths.

[0136] 3. Limited knowledge reasoning capabilities hinder intelligent decision-making support for complex faults. In actual power system operation and maintenance, it is often necessary to predict and warn of potential risks based on historical events, operational logic, and equipment dependencies. Existing graph technologies lack high-level reasoning capabilities, making it difficult to provide in-depth support for fault prediction, emergency response, and operation and maintenance scheduling. This application's embodiments combine the CIM model to extend the semantic relationships of power equipment, constructing a rule-driven graph structure reasoning framework that integrates equipment topology, operating status, and fault logic. This enables rapid matching of typical fault paths and identification of causal links, improving the graph's auxiliary decision-making capabilities in fault diagnosis and emergency response.

[0137] To address the aforementioned technical issues, this application proposes a plant fault knowledge graph modeling method that combines causal reasoning technology with graph structure evolution mechanism. It adopts a four-layer architecture of "data layer - modeling layer - evolution layer - service layer", integrating multi-source heterogeneous data parsing, causal relationship modeling, graph dynamic evolution and reasoning services to achieve dynamic modeling and root cause reasoning of complex plant faults.

[0138] Compared with the prior art, the plant fault knowledge graph modeling method proposed in this application, which combines causal reasoning technology with graph structure evolution mechanism, has the following significant advantages and positive effects.

[0139] 1. Introducing causal reasoning to improve the accuracy of root cause analysis. This application's embodiments, based on traditional knowledge graph construction, add a causal reasoning step. By explicitly modeling the causal chain of equipment state changes, the root cause of the fault can be accurately identified. This application's embodiments can overcome the shortcomings of existing technologies based on correlation analysis, not only discovering the connections between fault phenomena but also deeply analyzing the causal path of the fault, greatly improving the accuracy and interpretability of fault diagnosis.

[0140] 2. Graph Structure Evolution Mechanism Enhances Dynamic Adaptability and Timeliness. This application's embodiments introduce a graph structure evolution mechanism, enabling automatic updates to the knowledge graph structure based on real-time changes in plant operating status and the dynamic process of fault propagation. Compared to existing technologies that rely on static graph structures, graph structure evolution makes the knowledge graph more adaptable to complex fault scenarios, promptly reflecting the impact of fault propagation, equipment maintenance, replacement, and other factors on equipment relationships, thus improving the system's timeliness and accuracy in responding to plant status changes.

[0141] 3. Multi-level reasoning support for optimized fault prediction and emergency decision-making. This application integrates causal reasoning and graph reasoning mechanisms, supporting not only fault root cause localization based on current data but also complex multi-level fault prediction and propagation path reasoning. This application can effectively support the operation and maintenance decision-making system in providing intelligent auxiliary decision-making, identifying potential fault risks in advance, and conducting emergency responses. For example, the system can predict possible fault propagation paths based on the causal relationships of historical events and provide early warnings and operational suggestions to operation and maintenance personnel.

[0142] 4. Intelligent visualization enhances fault diagnosis and maintenance efficiency. This application's embodiments introduce intelligent and visual steps into the knowledge graph construction process. Through the visualized graph interface, maintenance personnel can intuitively see the causal relationships between devices and the fault evolution process. The graph not only possesses clear hierarchy and intuitive structure, but also dynamically displays the evolution of plant equipment status over time, providing maintenance personnel with real-time and accurate fault diagnosis basis, significantly improving fault analysis efficiency.

[0143] Figure 5 This is a schematic diagram of the structure of a substation fault handling device that combines causal reasoning knowledge graph modeling, as provided in an embodiment of this application.

[0144] like Figure 5 As shown, the substation fault handling device 5, which combines causal reasoning knowledge graph modeling, includes:

[0145] The first acquisition module 51 is used to acquire spatiotemporally aligned datasets by acquiring multi-source heterogeneous data and performing data cleaning processing. The spatiotemporally aligned datasets include one or more four-element data groups, which include device identifier, timestamp, feature vector, and event tag.

[0146] The second modeling module 52 is used to perform causal modeling processing on the spatiotemporal aligned dataset to obtain a causal graph, wherein the causal graph includes node information and relationship information between nodes.

[0147] The third construction module 53 is used to construct a spatiotemporal graph neural network model based on the causal graph and the device connection graph. The spatiotemporal graph neural network model realizes dynamic evolution of the graph based on an incremental update strategy.

[0148] The fourth service module 54 is used to output fault root cause location information based on the spatiotemporal graph neural network model.

[0149] Another embodiment of the present invention discloses a substation fault handling device 5 that combines causal reasoning knowledge graph modeling. This embodiment is based on the above... Figure 5 Based on the corresponding embodiments, the data cleaning process includes at least one of the following: invalid value processing, time alignment, feature normalization, and recoding.

[0150] Another embodiment of the present invention discloses a substation fault handling device 5 that combines causal reasoning knowledge graph modeling. This embodiment is based on the above... Figure 5 Based on the corresponding embodiment, the second modeling module 52 is used for:

[0151] A sliding window-based Granger causality detection method is used to perform causal analysis on the device runtime sequence data in the spatiotemporally aligned dataset to obtain the first causal information;

[0152] The first causal information is corrected based on the device connection relationship graph to obtain the causal graph.

[0153] Another embodiment of the present invention discloses a substation fault handling device 5 that combines causal reasoning knowledge graph modeling. This embodiment is based on the above... Figure 5 Based on the corresponding embodiment, the node information of the node includes at least one of the following: device identifier, fault event, and status indicator; the relationship information between the nodes includes at least one of the following: causal dependency relationship between nodes, causal strength, time delay, and confidence label; the causal strength is determined based on Granger coefficient factor, weight coefficient of Granger coefficient factor, Bayesian confidence factor, weight coefficient of Bayesian confidence factor, historical fault co-occurrence frequency factor, and weight coefficient of historical fault co-occurrence frequency factor, wherein the weight coefficient of Granger coefficient factor, the weight coefficient of Bayesian confidence factor, and the weight coefficient of historical fault co-occurrence frequency factor are updated based on time changes or system state changes.

[0154] Another embodiment of the present invention discloses a substation fault handling device 5 that combines causal reasoning knowledge graph modeling. This embodiment is based on the above... Figure 5 Based on the corresponding embodiments, the third construction module 53 is used for:

[0155] A knowledge graph ontology structure is constructed based on the device connection relationship graph. The knowledge graph ontology structure includes entity type, relationship type, and attribute information.

[0156] Based on the knowledge graph ontology structure and the causal graph, a spatiotemporal graph neural network model is constructed. The spatiotemporal graph neural network model implements an update strategy to achieve dynamic evolution of the graph when a new alarm event is generated or when the causal weight changes exceed a threshold.

[0157] Another embodiment of the present invention discloses a substation fault handling device 5 that combines causal reasoning knowledge graph modeling. This embodiment is based on the above... Figure 5 Based on the corresponding embodiments, the fourth service module 54 is used for at least one of the following:

[0158] The active alarm list is input into the spatiotemporal graph neural network model, which outputs multiple device identifiers as fault root cause location information.

[0159] The simulated scenario is input into the spatiotemporal graph neural network model, and a quantitative comparison report is output as the fault root cause location information. The simulated scenario includes any one of the following: simulated parameter correction scenario, simulated topology reconstruction scenario, simulated environmental intervention scenario, simulated protection logic optimization scenario, and simulated operation and maintenance measurement adjustment scenario.

[0160] When abnormal parameters are detected, warning information is output based on the spatiotemporal graph neural network model.

[0161] Another embodiment of the present invention discloses a substation fault handling device 5 that combines causal reasoning knowledge graph modeling. This embodiment is based on the above... Figure 5 Based on the corresponding embodiment, the fourth service module 54 is used for:

[0162] The causal topology of the knowledge graph and real-time device data are input into the spatiotemporal graph neural network model to obtain multiple risk paths, which include paths between multiple nodes.

[0163] A random walk algorithm with restart is used to evaluate the risk coefficients of the multiple risky paths and identify one or more high-risk paths.

[0164] The simulated scenario corresponding to one or more high-risk paths is input into the spatiotemporal graph neural network model, and the key causal path is output as the root cause location information of the fault and the spatiotemporal graph neural network model is optimized.

[0165] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0166] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0167] This application also provides an electronic device, such as... Figure 6 As shown, the electronic device 6 includes: at least one processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the at least one processor 61, wherein the processor 61 executes the computer program 63 to implement the steps in any of the above-described method embodiments.

[0168] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0169] This application provides a computer program product, including a computer program, which, when run, causes the steps in the above-described method embodiments to be executed.

[0170] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographic device / electronic device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0171] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0172] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0173] In the embodiments provided in this application, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings or direct couplings or communication connections may be through some interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0174] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0175] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A substation fault handling method combined with causal reasoning knowledge graph modeling, characterized in that, The method comprises the following steps: obtaining a spatio-temporal alignment data set by collecting multi-source heterogeneous data and performing data cleaning processing, the spatio-temporal alignment data set comprising one or more quadruple data groups, the quadruple data group comprising a device identifier, a timestamp, a feature vector, and an event label; performing causal modeling processing on the spatio-temporal alignment data set to obtain a causal graph, the causal graph comprising node information of nodes and relationship information between the nodes; constructing a spatio-temporal graph neural network model according to the causal graph and a device connection relationship graph, the spatio-temporal graph neural network model realizing dynamic evolution of a graph based on an incremental updating strategy; outputting fault root cause positioning information according to the spatio-temporal graph neural network model.

2. The method of claim 1, wherein, The data cleaning processing comprises at least one of the following: invalid value processing, time alignment, feature normalization, and recoding.

3. The method of claim 1, wherein, The step of performing causal modeling processing on the spatio-temporal alignment data set to obtain a causal graph comprises: performing causal analysis on device operation time series data in the spatio-temporal alignment data set by using a Granger causal detection method based on a sliding window to obtain first causal information; correcting the first causal information based on a device connection relationship graph to obtain a causal graph.

4. The method of claim 3, wherein, The node information of the nodes comprises at least one of the following: a device identifier, a fault event, and a state indicator; the relationship information between the nodes comprises at least one of the following: a causal dependency relationship between the nodes, a causal strength, a time delay, and a confidence label; the causal strength is determined based on a Granger coefficient factor, a weight coefficient of the Granger coefficient factor, a Bayesian confidence factor, a weight coefficient of the Bayesian confidence factor, a historical fault co-occurrence frequency factor, and a weight coefficient of the historical fault co-occurrence frequency factor, wherein the weight coefficient of the Granger coefficient factor, the weight coefficient of the Bayesian confidence factor, and the weight coefficient of the historical fault co-occurrence frequency factor are updated based on a change in time or a change in system state.

5. The method of claim 1, wherein, The step of constructing a spatio-temporal graph neural network model according to the causal graph and a device connection relationship graph, the spatio-temporal graph neural network model realizing dynamic evolution of a graph based on an incremental updating strategy, comprises: constructing a knowledge graph ontology structure according to a device connection relationship graph, the knowledge graph ontology structure comprising entity types, relationship types, and attribute information; constructing a spatio-temporal graph neural network model based on the knowledge graph ontology structure and the causal graph, the spatio-temporal graph neural network model realizing dynamic evolution of a graph by updating strategy under the condition of generating a new alarm event or when a causal weight change exceeds a threshold.

6. The method of claim 1, wherein, The step of outputting fault root cause positioning information according to the spatio-temporal graph neural network model comprises at least one of the following: inputting an active alarm list into the spatio-temporal graph neural network model to output a plurality of device identifiers as fault root cause positioning information; inputting a simulation scenario into the spatio-temporal graph neural network model to output a quantitative comparison report as fault root cause positioning information, the simulation scenario comprising any one of the following: a simulation parameter correction scenario, a simulation topology reconstruction scenario, a simulation environment intervention scenario, a simulation protection logic optimization scenario, and a simulation operation and maintenance measurement adjustment scenario. When an abnormal parameter is detected, early warning information is output based on the spatio-temporal graph neural network model.

7. The method of claim 1, wherein, The fault root cause positioning information is output according to the spatio-temporal graph neural network model, including: The causal topological structure of the knowledge graph and real-time device data are input into the spatio-temporal graph neural network model to obtain a plurality of risk paths, the risk paths including paths between a plurality of nodes; A random walk algorithm with restart is used to evaluate risk coefficients of the plurality of risk paths to determine one or more high-risk paths; The simulated scenarios corresponding to the one or more high-risk paths are input into the spatio-temporal graph neural network model to output a key causal path as the fault root cause positioning information and optimize the spatio-temporal graph neural network model.

Images